Towards definition extraction using conditional random fields

Abstract

Definition Extraction (DE) and terminology are contributing to help structuring the overwhelming amount of information available. This article presents KESSI (Knowledge Extraction System for Scientific/nInterviews), a multilingual domainindependent machine-learning approach to the extraction of definitional knowledge, specifically oriented to scientific interviews. The DE task was approached as both a classification and a sequential labelling task. In the latter, figures of Precision, Recall and F-Measure were similar to human annotation, and suggest that combining structural, statistical and linguistic/nfeatures with Conditional Random Fields can contribute significantly to the development of DE systems.